The Generalization Capability Investigation of CNN Model for Sleep Staging

The generalization capability of automated sleep staging is crucial for sleep disorder diagnosis. Convolutional Neural Networks (CNNs) can effectively extract staging features from physiological signals. This study developed a CNN-based automatic sleep staging model and evaluated its generalization performance. The model was trained using first-night recordings from 10 subjects in the Sleep-EDF-20 dataset, with generalization evaluated on follow-up recordings from the original cohort and two-night recordings from 9 new subjects. On the test dataset, the model demonstrated strong performance across all metrics, with an accuracy of 93.1%, a precision of 90.74%, a recall of 88.95%, and macro F1-score of $\mathbf{8 9. 6 8 \%}$. Moreover, the mean generalization capability scores obtained on an entirely novel generalization evaluation dataset were measured at 91.03%, $75.6 \%, 79.64 \%$, and 75.46%, respectively. The generalization analysis showed that wake stage classification achieved the highest recall (95.15%), followed by N3 (91.69%), N2 (89.46%), and REM (73.24%). Notably, N1 recall (48.65%) underperformed the testing dataset baseline (67.61%) by 28%. The performance deterioration in N1 stage classification revealed that conventional holdout dataset validation alone was insufficient to ensure generalizability for this particular sleep stage.